Image Classification And Annotation Based On Probabilistic Graphical Model

In image classification and annotation, image features and learning methods are used to categorize and annotate images.Although these tasks are usually easy and natural for human beings, it is difficult for computers. Despite the great advance of image classification and annotation techniques in the past few years, it is still one of the fundamental challenges in computer vision. The purposes of this paper are as follow:First, improving sLDA model with different kinds of response variables; Second embeding the prososed model of image annotation into the multi-class sLDA; Third, applying the above results to LabelMe imag database.We will focus on two main types of works for image classification and annotation task. One type can be viewed as bag-of-word model and another is part-based model. First, this thesis describes the classical topic models in detail:(1) Naive Bayes model; (2) PLSI model; (3) LDA model; (4) sLDA model; Then, we propose a new model based on sLDA, which originally was designed for real-valued reponse by normal linear model. Recently multi-class sLDA for image classification improved sLDA to handle discrete class label response variable. In this thesis, the multi-variate binary response of image annotation data comes from a logistic linear model by logistic function. We embed the prososed model of image annotation into the multi-class sLDA. This yields a single coherent model of images, class labels and annotation terms, allowing classification and annotation to be performed simultaneously. We derive a maximum-likelihood procedure for parameter estimation based on mean-field variational approximations; Finally, We test the improved model on real-world image date sets for image classification and annotation, resulting that in classification the new model reduce the error of Fei-fei and Perona by 12% on LabelMe dataset, increase performance comparing to multi-class sLDA correspondingly, in annotation the new model improve caption prediction probability than cLDA by 6%...